fully resetting the slot state if we are in a prefilling state on an …

…unexpected call to get_num_new_matched_tokens

lib/bindings/python/rust/llm/block_manager/vllm/connector/leader.rs

@@ -139,6 +139,11 @@ impl Leader for KvConnectorLeader {
             .lock()
             .map_err(|e| anyhow::anyhow!("Failed to lock slot: {}", e))?;
 
+        if slot.state() == SlotState::Prefilling {
+            tracing::warn!("slot is in the Prefilled state; this seems like we need to reset the slot and start over");
+            slot.reset();
+        }
+
         // early exit if we cannot match full block
         if (slot.sequence().total_tokens() - num_computed_tokens) < self.block_size {
             return Ok((0, false));
@@ -319,7 +324,7 @@ impl Leader for KvConnectorLeader {
 
                 // todo: we probably need to reset the slot state and reload it from `cache_req`; however, we do not
                 // know if it will take another pass at `get_num_new_matched_tokens` or `update_state_after_alloc`.
-                slot.reset_after_preemption()?;
+                slot.reset_after_preemption();
 
                 // note, we can not trigger onboarding here -- perhaps we are supposed to or perhaps will get another
                 // pass at `get_num_new_matched_tokens` or `update_state_after_alloc`.

lib/bindings/python/rust/llm/block_manager/vllm/connector/leader/slot.rs

@@ -127,7 +127,10 @@ pub trait Slot: std::fmt::Debug {
     fn record_cached_disk_tokens(&mut self, num_tokens: usize);
 
     /// Reset the slot after preemption.
-    fn reset_after_preemption(&mut self) -> Result<(), SlotError>;
+    fn reset_after_preemption(&mut self);
+
+    /// Reset the slot.
+    fn reset(&mut self);
 }
 
 pub trait ExternallyManagedDeviceSlot: Slot {
@@ -347,7 +350,7 @@ impl Slot for VllmConnectorSlot {
         self.state
     }
 
-    fn reset_after_preemption(&mut self) -> Result<(), SlotError> {
+    fn reset_after_preemption(&mut self) {
         assert!(self.staging_from_disk.is_none());
         assert!(self.staging_from_host.is_none());
         assert!(self.pending_operations.is_none());
@@ -360,7 +363,11 @@ impl Slot for VllmConnectorSlot {
         self.tokens_cached_from_device = 0;
         self.tokens_cached_from_host = 0;
         self.tokens_cached_from_disk = 0;
-        Ok(())
+    }
+
+    fn reset(&mut self) {
+        self.reset_after_preemption();
+        self.state = SlotState::Initialized;
     }
 
     fn record_cached_device_tokens(&mut self, num_tokens: usize) {

lib/llm/src/lib.rs

@@ -38,6 +38,9 @@ pub mod types;
 #[cfg(feature = "block-manager")]
 pub mod block_manager;
 
+#[cfg(feature = "block-manager")]
+pub mod integrations;
+
 /// Reads a JSON file, extracts a specific field, and deserializes it into type T.
 ///
 /// # Arguments